From c9a8c526acd185176e486bee4624039740f8c435 Mon Sep 17 00:00:00 2001 From: Andrew Senkevich Date: Thu, 18 Jun 2015 17:55:55 +0300 Subject: Vector sincos for x86_64 and tests. Here is implementation of vectorized sincos containing SSE, AVX, AVX2 and AVX512 versions according to Vector ABI . * NEWS: Mention addition of x86_64 vector sincos. * bits/libm-simd-decl-stubs.h: Added stubs for sincos. * math/math.h (__MATHDECL_VEC): New macro. * math/bits/mathcalls.h: Added sincos declaration with __MATHDECL_VEC. * math/gen-libm-have-vector-test.sh: Added generation of sincos wrapper declaration under condition. * math/test-vec-loop.h (TEST_VEC_LOOP): Refactored. * math/test-double-vlen2.h: Added wrapper for sincos tests, reflected TEST_VEC_LOOP change. * math/test-double-vlen4.h: Likewise. * math/test-double-vlen8.h: Likewise. * math/test-float-vlen16.h: Reflected TEST_VEC_LOOP change. * math/test-float-vlen4.h: Likewise. * math/test-float-vlen8.h: Likewise. * sysdeps/unix/sysv/linux/x86_64/libmvec.abilist: New symbols added. * sysdeps/x86/fpu/bits/math-vector.h: Added sincos SIMD declaration. * sysdeps/x86_64/fpu/Makefile (libmvec-support): Added new files. * sysdeps/x86_64/fpu/Versions: New versions added. * sysdeps/x86_64/fpu/libm-test-ulps: Regenerated. * sysdeps/x86_64/fpu/multiarch/Makefile (libmvec-sysdep_routines): Added build of SSE, AVX2 and AVX512 IFUNC versions. * sysdeps/x86_64/fpu/multiarch/svml_d_sincos2_core.S: New file. * sysdeps/x86_64/fpu/multiarch/svml_d_sincos2_core_sse4.S: New file. * sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core.S: New file. * sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core_avx2.S: New file. * sysdeps/x86_64/fpu/multiarch/svml_d_sincos8_core.S: New file. * sysdeps/x86_64/fpu/multiarch/svml_d_sincos8_core_avx512.S: New file. * sysdeps/x86_64/fpu/svml_d_sincos2_core.S: New file. * sysdeps/x86_64/fpu/svml_d_sincos4_core.S: New file. * sysdeps/x86_64/fpu/svml_d_sincos4_core_avx.S: New file. * sysdeps/x86_64/fpu/svml_d_sincos8_core.S: New file. * sysdeps/x86_64/fpu/svml_d_sincos_data.S: New file. * sysdeps/x86_64/fpu/svml_d_sincos_data.h: New file. * sysdeps/x86_64/fpu/svml_d_wrapper_impl.h: Added wrappers for sincos. * sysdeps/x86_64/fpu/test-double-vlen2-wrappers.c: Vector sincos tests. * sysdeps/x86_64/fpu/test-double-vlen2.c: Likewise. * sysdeps/x86_64/fpu/test-double-vlen4-avx2-wrappers.c: Likewise. * sysdeps/x86_64/fpu/test-double-vlen4-avx2.c: Likewise. * sysdeps/x86_64/fpu/test-double-vlen4-wrappers.c: Likewise. * sysdeps/x86_64/fpu/test-double-vlen4.c: Likewise. * sysdeps/x86_64/fpu/test-double-vlen8-wrappers.c: Likewise. * sysdeps/x86_64/fpu/test-double-vlen8.c: Likewise. --- .../fpu/multiarch/svml_d_sincos4_core_avx2.S | 277 +++++++++++++++++++++ 1 file changed, 277 insertions(+) create mode 100644 sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core_avx2.S (limited to 'sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core_avx2.S') diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core_avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core_avx2.S new file mode 100644 index 0000000000..ec1ccc6357 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_sincos4_core_avx2.S @@ -0,0 +1,277 @@ +/* Function sincos vectorized with AVX2. + Copyright (C) 2014-2015 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + . */ + +#include +#include "svml_d_sincos_data.h" + + .text +ENTRY (_ZGVdN4vvv_sincos_avx2) +/* + ALGORITHM DESCRIPTION: + + ( low accuracy ( < 4ulp ) or enhanced performance + ( half of correct mantissa ) implementation ) + + Argument representation: + arg = N*Pi + R + + Result calculation: + sin(arg) = sin(N*Pi + R) = (-1)^N * sin(R) + arg + Pi/2 = (N'*Pi + R') + cos(arg) = sin(arg+Pi/2) = sin(N'*Pi + R') = (-1)^N' * sin(R') + sin(R), sin(R') are approximated by corresponding polynomial. */ + + pushq %rbp + cfi_adjust_cfa_offset (8) + cfi_rel_offset (%rbp, 0) + movq %rsp, %rbp + cfi_def_cfa_register (%rbp) + andq $-64, %rsp + subq $448, %rsp + movq __svml_dsincos_data@GOTPCREL(%rip), %rax + vmovups %ymm14, 288(%rsp) + vmovups %ymm8, 352(%rsp) + vmovupd __dSignMask(%rax), %ymm6 + vmovupd __dInvPI(%rax), %ymm2 + vmovupd __dPI1_FMA(%rax), %ymm5 + vmovups %ymm9, 224(%rsp) + +/* ARGUMENT RANGE REDUCTION: + Absolute argument: X' = |X| */ + vandnpd %ymm0, %ymm6, %ymm1 + +/* SinY = X'*InvPi + RS : right shifter add */ + vfmadd213pd __dRShifter(%rax), %ymm1, %ymm2 + +/* SinSignRes = Y<<63 : shift LSB to MSB place for result sign */ + vpsllq $63, %ymm2, %ymm4 + +/* SinN = Y - RS : right shifter sub */ + vsubpd __dRShifter(%rax), %ymm2, %ymm2 + +/* SinR = X' - SinN*Pi1 */ + vmovdqa %ymm1, %ymm14 + vfnmadd231pd %ymm2, %ymm5, %ymm14 + +/* SinR = SinR - SinN*Pi1 */ + vfnmadd231pd __dPI2_FMA(%rax), %ymm2, %ymm14 + +/* Sine result sign: SinRSign = SignMask & SinR */ + vandpd %ymm14, %ymm6, %ymm7 + +/* Set SinRSign to 0.5 */ + vorpd __dOneHalf(%rax), %ymm7, %ymm3 + +/* CosN = SinN +(-)0.5 */ + vaddpd %ymm3, %ymm2, %ymm3 + +/* CosR = SinX - CosN*Pi1 */ + vmovdqa %ymm1, %ymm8 + vfnmadd231pd %ymm3, %ymm5, %ymm8 + vmovupd __dPI3_FMA(%rax), %ymm5 + vcmpnle_uqpd __dRangeVal(%rax), %ymm1, %ymm1 + +/* CosR = CosR - CosN*Pi2 */ + vfnmadd231pd __dPI2_FMA(%rax), %ymm3, %ymm8 + +/* SinR = SinR - SinN*Pi3 */ + vfnmadd213pd %ymm14, %ymm5, %ymm2 + +/* CosR = CosR - CosN*Pi3 */ + vfnmadd213pd %ymm8, %ymm5, %ymm3 + vmovupd __dC6(%rax), %ymm8 + +/* SinR2 = SinR^2 */ + vmulpd %ymm2, %ymm2, %ymm14 + +/* CosR2 = CosR^2 */ + vmulpd %ymm3, %ymm3, %ymm5 + +/* Grab SignX */ + vandpd %ymm0, %ymm6, %ymm9 + +/* Update CosRSign and CosSignRes signs */ + vxorpd %ymm6, %ymm7, %ymm6 + vxorpd %ymm6, %ymm4, %ymm7 + +/* Update sign SinSignRes */ + vxorpd %ymm9, %ymm4, %ymm6 + +/* Polynomial approximation */ + vmovupd __dC7(%rax), %ymm4 + vmovdqa %ymm8, %ymm9 + vfmadd231pd __dC7(%rax), %ymm14, %ymm9 + vfmadd213pd %ymm8, %ymm5, %ymm4 + vfmadd213pd __dC5(%rax), %ymm14, %ymm9 + vfmadd213pd __dC5(%rax), %ymm5, %ymm4 + vfmadd213pd __dC4(%rax), %ymm14, %ymm9 + vfmadd213pd __dC4(%rax), %ymm5, %ymm4 + +/* SinPoly = C3 + SinR2*(C4 + SinR2*(C5 + SinR2*(C6 + SinR2*C7))) */ + vfmadd213pd __dC3(%rax), %ymm14, %ymm9 + +/* CosPoly = C3 + CosR2*(C4 + CosR2*(C5 + CosR2*(C6 + CosR2*C7))) */ + vfmadd213pd __dC3(%rax), %ymm5, %ymm4 + +/* SinPoly = C2 + SinR2*SinPoly */ + vfmadd213pd __dC2(%rax), %ymm14, %ymm9 + +/* CosPoly = C2 + CosR2*CosPoly */ + vfmadd213pd __dC2(%rax), %ymm5, %ymm4 + +/* SinPoly = C1 + SinR2*SinPoly */ + vfmadd213pd __dC1(%rax), %ymm14, %ymm9 + +/* CosPoly = C1 + CosR2*CosPoly */ + vfmadd213pd __dC1(%rax), %ymm5, %ymm4 + +/* SinPoly = SinR2*SinPoly */ + vmulpd %ymm14, %ymm9, %ymm8 + +/* CosPoly = CosR2*CosPoly */ + vmulpd %ymm5, %ymm4, %ymm4 + +/* SinPoly = SinR*SinPoly */ + vfmadd213pd %ymm2, %ymm2, %ymm8 + +/* CosPoly = CosR*CosPoly */ + vfmadd213pd %ymm3, %ymm3, %ymm4 + vmovmskpd %ymm1, %ecx + +/* Final reconstruction + Update Sin result's sign */ + vxorpd %ymm6, %ymm8, %ymm3 + +/* Update Cos result's sign */ + vxorpd %ymm7, %ymm4, %ymm2 + testl %ecx, %ecx + jne .LBL_1_3 + +.LBL_1_2: + cfi_remember_state + vmovups 352(%rsp), %ymm8 + vmovups 224(%rsp), %ymm9 + vmovups 288(%rsp), %ymm14 + vmovupd %ymm2, (%rsi) + vmovdqa %ymm3, (%rdi) + movq %rbp, %rsp + cfi_def_cfa_register (%rsp) + popq %rbp + cfi_adjust_cfa_offset (-8) + cfi_restore (%rbp) + ret + +.LBL_1_3: + cfi_restore_state + vmovupd %ymm0, 256(%rsp) + vmovupd %ymm3, 320(%rsp) + vmovupd %ymm2, 384(%rsp) + je .LBL_1_2 + + xorb %dl, %dl + xorl %eax, %eax + vmovups %ymm10, 128(%rsp) + vmovups %ymm11, 96(%rsp) + vmovups %ymm12, 64(%rsp) + vmovups %ymm13, 32(%rsp) + vmovups %ymm15, (%rsp) + movq %rsi, 160(%rsp) + movq %r12, 200(%rsp) + cfi_offset_rel_rsp (12, 200) + movb %dl, %r12b + movq %r13, 192(%rsp) + cfi_offset_rel_rsp (13, 192) + movl %eax, %r13d + movq %r14, 184(%rsp) + cfi_offset_rel_rsp (14, 184) + movl %ecx, %r14d + movq %r15, 176(%rsp) + cfi_offset_rel_rsp (15, 176) + movq %rbx, 168(%rsp) + movq %rdi, %rbx + cfi_remember_state + +.LBL_1_6: + btl %r13d, %r14d + jc .LBL_1_13 + +.LBL_1_7: + lea 1(%r13), %esi + btl %esi, %r14d + jc .LBL_1_10 + +.LBL_1_8: + incb %r12b + addl $2, %r13d + cmpb $16, %r12b + jb .LBL_1_6 + + vmovups 128(%rsp), %ymm10 + movq %rbx, %rdi + vmovups 96(%rsp), %ymm11 + vmovups 64(%rsp), %ymm12 + vmovups 32(%rsp), %ymm13 + vmovups (%rsp), %ymm15 + vmovupd 320(%rsp), %ymm3 + vmovupd 384(%rsp), %ymm2 + movq 160(%rsp), %rsi + movq 200(%rsp), %r12 + cfi_restore (%r12) + movq 192(%rsp), %r13 + cfi_restore (%r13) + movq 184(%rsp), %r14 + cfi_restore (%r14) + movq 176(%rsp), %r15 + cfi_restore (%r15) + movq 168(%rsp), %rbx + jmp .LBL_1_2 + +.LBL_1_10: + cfi_restore_state + movzbl %r12b, %r15d + shlq $4, %r15 + vmovsd 264(%rsp,%r15), %xmm0 + vzeroupper + + call sin@PLT + + vmovsd %xmm0, 328(%rsp,%r15) + vmovsd 264(%rsp,%r15), %xmm0 + + call cos@PLT + + vmovsd %xmm0, 392(%rsp,%r15) + jmp .LBL_1_8 + +.LBL_1_13: + movzbl %r12b, %r15d + shlq $4, %r15 + vmovsd 256(%rsp,%r15), %xmm0 + vzeroupper + + call sin@PLT + + vmovsd %xmm0, 320(%rsp,%r15) + vmovsd 256(%rsp,%r15), %xmm0 + + call cos@PLT + + vmovsd %xmm0, 384(%rsp,%r15) + jmp .LBL_1_7 + +END (_ZGVdN4vvv_sincos_avx2) -- cgit 1.4.1